【作者】 谢桐瑜；

【导师】 许少华；

【作者基本信息】 大庆石油学院 ， 计算机应用技术， 2008， 硕士

【摘要】 时间序列预测是用被预测事物过去和现在的观测数据,构造依时间变化的序列模型,借助其反映出来的发展过程、方向和趋势,进行外推或延伸,从而预测下一段时间可能达到的状态和水平,是预测领域的重要组成部分,在工程领域和科学研究中具有重要意义。过程神经元网络可以同时表达系统多输入时变信号的空间聚合作用和时间过程的累积效应,能够直接处理过程式数据,且具有在时变函数空间上的连续性和对满足Lipschitz条件泛函的逼近能力,以及与Turing机等价的计算能力,这使过程神经元网络在解决实际问题中有着广泛的适应性和灵活性。论文对已有的用于时序预测问题研究的理论、方法、模型进行归纳总结,对目前方法存在的问题和困难进行了分析,并基于时序预测模型及径向基过程神经网络,提出了一种新的非线性组合预测方法。该方法将单一预测方法得出的预测结果,作为RBFPNN的输入,而实际的历史数据值作为网络的期望输出,这样避免了一般线性组合预测方法中确定各个权重的复杂计算,又可涵盖实际问题的线性与非线性两方面,综合地利用各种单一预测方法提供的信息,提高了预测精度。最后将这种预测方法应用在航空公司的旅客数量预测中,取得了满意的结果。针对非线性时间序列中的混沌时间序列分析问题,论文从重构相空间理论出发,探讨确定相空间嵌入维数和延迟时间各种不同的方法。对时间序列混沌特性的识别方法,及混沌相空间预测模型进行了详细讨论,并根据相空间重构技术和过程神经网络的技术原理,寻求二者的结合点,从而提出了一种相空间重构与过程神经网络相结合的混沌时间序列预测方法。并以太阳黑子数预测为例验证了算法的有效性。针对时序预测中的Markov链状态转移预测问题,提出了一种基于离散过程神经元网络(DPNN)的等效状态转移预测方法和模型,探讨了DPNN与Markov模型在一定条件下对于系统状态转移问题描述的等价关系。对于任意的Markov链,本文给出了与之等效DPNN的构建方法和Markov链状态转移概率条件约束下的网络权值矩阵求解算法,仿真实验结果验证了方法的有效性。更多还原

【Abstract】 Time series prediction is forecast to be things of the past and present observational data, structural changes in the time-series model and use it to reflect in the development process, the direction and trends, extrapolation or extended, which is expected to predict a period of time , the forecast is an important component of the field, in engineering and scientific research. Process neural networks expression system can be more time-varying input signals in the polymerization of space and time course of the cumulative effect, it can directly process data, and has a time-varying function in the continuity of space and to meet Lipschitz condition functional approaching capacity, as well as the Turing machine equivalent computing power, which makes the process neural network in the solution of practical problems of a wide range of adaptability and flexibility.This paper has been forecast for timing problems with the theories, methods, models summary, the present method of the problems and difficulties of the analysis, and forecasting model based on the timing and process of RBF neural network, a combination of new nonlinear forecasting methods. The method means that the forecast results of a single forecast, as RBFPNN input, the actual value of the historical data network as the expectations of output, so to avoid a general linear combination of forecasting methods identified in the various weight of the complex, could cover the practical problems of both linear and non-linear, integrated use of a single prediction method to provide the information and improve the prediction accuracy. Finally, the methods used in this projection airline forecast in the number of passengers, and achieved satisfactory results.Nonlinear time series against the chaotic time series, the paper from the reconstruction phase space theory, determine the phase space of embedding dimension and a variety of time delay methods. Chaos on the time sequence of the identification method, and chaotic phase space forecasting model were discussed in detail, and in accordance with phase space reconstruction process neural network technology and the technical principles for the combination of the two, and thus proposed a phase-space neural network structure and the process of combining chaotic time series forecasting methods. And the sunspot forecasting as an example verifies the effectiveness of the algorithm.Against the prediction of the state Markov chain prediction transfer, a process based on discrete neural network (DPNN) transfer the equivalent state forecasting methods and models were created, this paper explored the DPNN Markov model under certain conditions for the transfer of state system Description of equivalence relations. For arbitrary Markov chain, this paper proposed the equivalent DPNN Construction methods and state Markov chain transition probability under the conditions of the network-weight-matrix algorithm, the simulation results demonstrate the effectiveness of the method.更多还原